Podcast Video to Three Vertical Clips With Captions: $0.66

The short answer

Three 45-second clips from a podcast recording cost $0.66: three POST /v1/video-trim jobs at $0.02 each ($0.06) and three standalone caption jobs at $0.20 each ($0.60), under the 60-second ceiling for the caption price. The source must be a Sume-hosted clip no longer than 1,800 seconds, which is 30 minutes.

Find the clips

Run video inspect with transcribe: true to get word timings; the transcript rate is $0.01 per audio minute, and the transcript request reserves a 1-minute hint unless you pass duration_seconds (maximum hint 600 seconds). For a 30-minute episode, run it on the passage you are searching, not the whole file. segmentation.mode: sentence returns sentence segments you can read as candidate cut points.

Pick start and end times at sentence edges. Video trim takes start and one of end or duration; a clip may last 0.2 to 900 seconds. Use precision: exact for a frame-accurate cut; keyframe copies the stream and can start up to one group of pictures early.

Make it vertical

A trim's output field can conform width, height and fps, but docs describe it as a conform step, not a reframing tool. For a vertical clip with a speaker in the center, use a crop in video filter ($0.02), whose crop is a fractional rectangle of the source frame, or lay a title card over the video with timeline compose ($0.02). Both run before captions.

Captions come last. Pick a style such as slam, punch or tiktok-green; for Korean speech pick a Hangul style, because Latin styles refuse Hangul text with a 400. Send source_caption_id to restyle without paying for speech-to-text again.

If a clip is longer than a minute, the standalone caption price in the docs is stated for videos of up to 60 seconds, so check the live price in GET /v1/catalog before you caption a 90-second clip. Keep your cuts under a minute and the $0.66 holds. Run dry_run or generation_admission_preview if you submit these through MCP. Give each trim its own Idempotency-Key, such as the episode id plus the start second, so a retry after a timeout cannot cut and bill the same clip twice. The trim result gives actual_start_seconds, which matters if you used keyframe precision and need to re-base the caption times.

• Silent clips fail with caption_no_speech; use cues for authored text.
• Get permission from every guest before you publish a clip of them.
• Check names and numbers in the burned captions; fix with words if needed.